Discussion Ryzen 3000 series benchmark thread ** Open **

Page 27 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Dec 6, 2018
118
90
61
So for the rest of us who don't care about the idle power savings and want a 1ms polling rate, we are now screwed cause people were complaining about a software layered observing app?
for the "Ryzen Balanced" profile
I don't think that is true, the 15ms is only when it wakes up not when you are busy using it, then it's is still 1ms.
 
Nov 26, 2005
14,396
25
91
Last edited:

B-Riz

Golden Member
Feb 15, 2011
1,223
300
136
Hmm might have to try that myself. 3800x might be binned slightly better than the 3900x!
I would consider the 3900X one 3600X and one 3600 combined, worst case, and two 3600X combined best case.
 
Apr 27, 2000
12,759
1,563
126
I would consider the 3900X one 3600X and one 3600 combined, worst case, and two 3600X combined best case.
Makes me wonder which dice AMD will use for the 3950X. If it's two 3800x dice . . . look out.

edit: also, I tried my hand at AIDA64's FPU stress test. @ 1.3v + Medium LLC, my chip will do 4400 @ 72C. 4050 is a non-starter. I have no desire to push higher than that at this time. The voltage setting is, according to CPU-z, accurate, though I haven't done multimeter tests to see if I'm still getting significant vdroop with Medium LLC. It's possible that I am (HWiNFO64 says so).
 
Last edited:
Nov 26, 2005
14,396
25
91
Last edited:

maddogmcgee

Senior member
Apr 20, 2015
223
59
101
Is the issue with the Windows Timer Resolution, Windows CPU Power configuration, or "Chipset Drivers" according to this article? Chipset Drivers to me has always meant the motherboard drivers for things like the NB, SB, etc. https://www.techpowerup.com/review/new-amd-chipset-driver-performance-test-ryzen-9-3900x/
The chipset driver changes the AMD balanced profile to poll every 15ms, reducing the higher voltage and temps my 3700x were getting with minimal stuff going on (and monitoring software hitting the cpu every 1ms). The high performance still does 1ms so if you use that you haven't lost any performance.
 
Nov 26, 2005
14,396
25
91
The chipset driver changes the AMD balanced profile to poll every 15ms, reducing the higher voltage and temps my 3700x were getting with minimal stuff going on (and monitoring software hitting the cpu every 1ms). The high performance still does 1ms so if you use that you haven't lost any performance.
That's basically what the default Balanced profile does for Intel Chips. My old i7 970 & Windows 10.

So AMD made a Processor Power Management change via the Power plan through their AMD Chipset drivers for Ryzen 3? Is there that big a latency with dual ccx chips or whatever they are called, ?
 

maddogmcgee

Senior member
Apr 20, 2015
223
59
101
That's basically what the default Balanced profile does for Intel Chips. My old i7 970 & Windows 10.

So AMD made a Processor Power Management change via the Power plan through their AMD Chipset drivers for Ryzen 3? Is there that big a latency with dual ccx chips or whatever they are called, ?
I don't think it was an issue with latency off the ccx's. the CPU's are designed to change voltage way faster than Intel CPU's and quickly boost to a high voltage/clock to provide snappy performance. The issue is some software was causing this behaviour to constantly happen needlessly. The AMD solution was to change the polling to 15ms. I don't think most of us will notice any difference. I am far from an expert though and am just going off the little info brochure AMD released a few days ago. I would link it but can't find it at the moment.
 
Nov 26, 2005
14,396
25
91
I don't think it was an issue with latency off the ccx's. the CPU's are designed to change voltage way faster than Intel CPU's and quickly boost to a high voltage/clock to provide snappy performance. The issue is some software was causing this behaviour to constantly happen needlessly. The AMD solution was to change the polling to 15ms. I don't think most of us will notice any difference. I am far from an expert though and am just going off the little info brochure AMD released a few days ago. I would link it but can't find it at the moment.
I think I seen the PDF around here somewhere explaining some things about the chipset drivers, or it might of been on the Overclock forums.
 
Dec 6, 2018
118
90
61
The binning here is crystal clear:

We are talking about +0.3GHz (less than 10%) difference between their best and worst dies.
If we assume in best case this is their bottom half of silicon quality then the best is +0.6GHz, still not really 5GHz.
I think the wrong assumption of 5GHz is that it's only valid for the same Zen die design and since Zen2 is bigger with faster IF you obviously cannot just expect the same clock gain. (ask intel with Sunny Cove)
On top of that we don't get any power gain of the I/O die circuits even if that wouldn't make that much difference on 7nm.
 

Hitman928

Golden Member
Apr 15, 2012
1,892
334
136

tl;dw - You can get a bit of a boost manually tweaking RAM timings compared to XMP settings. A couple games showed the biggest improvement.

e.g. Far Cry New Dawn

1565188240762.png
 

Elfear

Diamond Member
May 30, 2004
6,855
90
126

tl;dw - You can get a bit of a boost manually tweaking RAM timings compared to XMP settings. A couple games showed the biggest improvement.

e.g. Far Cry New Dawn

View attachment 9315
:openmouth: Those are some impressive gains! I wonder if Intel gets a similar boost with tweaked sub-timings? IIRC, when a review site tested Gen 1 Ryzen and Coffeelake with XMP and optimized sub-timings, Intel saw increased performance but not as much as AMD got.
 
Apr 27, 2000
12,759
1,563
126
Shoot I've been tweaking subtimings since 2009! Only issue I'm having on my 3900x is that if you get too close to the razor's edge, sometimes you lose performance which is really weird.
 
Last edited:
Sep 4, 2016
88
53
61
I see the same behavior with 1st gen Ryzen and B-Die memory... Albeit stable some tight timing combinations actually regress in performance.

That's why tuning is time consuming: you need to test more than Aida64 to get the real performance impact and that's for each small change...
 
Apr 27, 2000
12,759
1,563
126
I see the same behavior with 1st gen Ryzen and B-Die memory... Albeit stable some tight timing combinations actually regress in performance.
My 1800x would just crash if I tightened the wrong timings or pushed for too much clockspeed. My 3900x will usually do that, but sometimes not so much. And performance regression isn't uniform. Something that runs "just fine" in y-cruncher might lead to regression in CBR20 (oddly enough).
 

Similar threads



ASK THE COMMUNITY

TRENDING THREADS